Flame stability monitoring method based on image analysis and data fusion

Abstract

The application discloses a flame stability monitoring method based on image analysis and data fusion, three types of flame images under different load combustion conditions are collected, a flame front band in the flame image is extracted from a self-luminous CH* flame image, a main combustion domain in the flame image is extracted from a self-luminous OH* flame image, the flame front band and the flame image form a closed unburned domain, the main combustion domain and the flame image are subtracted to obtain a burn-out domain, three domains and one band are obtained, corresponding parameters are extracted, and statistical analysis is conducted on the parameters to distinguish flame stability and flame instability. The application relies on the flame image to identify the flame stability, utilizes a comprehensive judgment flame stability index RSI, and realizes quantitative identification of the flame stability. Compared with identification of the flame stability only by means of the high-speed flame image, the flame stability identification precision is improved by 34%. Compared with a flame stability index relying on a shallow feature, the identification precision is improved by 23%, the monitoring precision is high, and the robustness of overcoming random interference is achieved.

Description

Technical Field

[0001] This invention relates to the field of flame stability monitoring, and in particular to a flame stability monitoring method based on image analysis and data fusion. Background Technology

[0002] Currently, carbon emissions from coal-fired boilers account for the majority of China's total carbon emissions. With the introduction of the "dual carbon" target, the newly installed capacity of coal-fired boilers has decreased year by year, while the newly installed capacity of non-fossil energy has increased year by year. However, due to the seasonal depletion of water resources and the intermittent and fluctuating nature of wind and solar energy resources, curtailment of solar and wind power has occurred. To reduce this curtailment, coal-fired boilers still need to play a crucial role in ensuring energy supply.

[0003] In the power generation industry, maintaining a stable combustion state over a long period is a fundamental requirement for safe operation. In actual operation, especially due to factors such as peak shaving, the frequency of combustion instability is increased. Combustion instability not only reduces combustion efficiency and increases NOx emissions, but in extreme cases, it can also cause furnace flameout or even explosions, posing a threat to the overall operational safety of the unit. Therefore, establishing an accurate and effective flame stability monitoring system is of significant practical importance for achieving optimized combustion control.

[0004] Currently, there are many methods for combustion monitoring in industry, mainly including optical combustion detectors and combustion detection systems. Optical combustion detectors utilize ultraviolet, visible, and infrared sensors to determine the combustion state by detecting the brightness, amplitude, and flicker frequency of the target combustion. However, these detectors are easily affected by factors such as load fluctuations, leading to inaccurate location identification and frequent false alarms and missed alarms. Combustion detection systems employ CCD camera technology, which can visually display the combustion area; however, the interpretation of flame images relies on operator experience and lacks objectivity. Self-luminescent groups, on the other hand, emit fluorescence when their excited state transitions to the ground or low-energy state, effectively displaying combustion chemical reactions in the ultraviolet to visible light range. These free groups, such as OH*, CH*, etc., can effectively display combustion chemical reactions. and These parameters can characterize aspects such as flame ignition, flame reaction zone, flame structure, heat release rate, and flame temperature. Therefore, self-emissive image analysis can be used as a method for assessing flame stability during combustion.

[0005] With the development of image processing technology, new detection methods for combustion diagnosis are constantly emerging. These methods typically utilize CCD cameras to acquire flame images, process the data, and output the detection results in real time. This not only inherits the intuitiveness of combustion television but also fully leverages the powerful data processing capabilities of computers, resulting in a qualitative improvement in flame image detection functionality. Currently, common data processing methods are mainly divided into two categories: one is frequency domain methods, which analyze data such as image brightness, combustion area, and power spectral density over a period of time for fault diagnosis. However, this method has a long response time and cannot reflect the current combustion state in real time. The other is supervised machine learning methods, which utilize shallow features of flame images to construct nonlinear monitoring models. However, these models are based on shallow features in flame images and prior human knowledge, resulting in less than ideal diagnostic performance, poor robustness, and the need for extensive labeled data for model training. Furthermore, most current data analysis methods can only determine the presence or absence of combustion but cannot quantitatively monitor flame stability, lacking guidance for boiler load adjustment.

[0006] Autoencoder networks (ANNs) are a type of unsupervised machine learning algorithm that can extract deep features from data in an unsupervised manner. Therefore, we can combine ANNs with traditional frequency domain methods and shallow feature extraction methods from machine learning to qualitatively and quantitatively identify the stability of flame images in real time, based on the composition of self-emitting radicals and flame images. Compared to past frequency domain methods and machine learning methods, ANNs can be pre-trained based on prior knowledge to obtain a more accurate model of flame stability in real time. Furthermore, the model can be updated iteratively by manually adding data to the flame image model based on actual boiler malfunctions. Summary of the Invention

[0007] To address the above problems, this invention proposes a flame stability monitoring method based on image analysis and data fusion. The specific technical solution is as follows:

[0008] A flame stability monitoring method based on image analysis and data fusion includes the following steps:

[0009] Step 1: Collect three types of flame images under different load combustion conditions: self-luminous OH* flame image, CH* flame image and high-speed flame image. Among them, the self-luminous OH* and CH* flame images are preprocessed to obtain flame image one, and the high-speed flame image is preprocessed to obtain flame image two.

[0010] Step 2: Extract the flame front band from the self-luminous CH* flame diagram and the main combustion zone from the self-luminous OH* flame diagram, respectively;

[0011] Step 3: The flame front beacon zone and the second flame image form a closed unburned region, and the main combustion region of the flame is subtracted from the first flame image to obtain the burnout region;

[0012] Step 4: Extract the thermodynamic parameters, geometric parameters, composition parameters, and frequency domain parameters corresponding to the main combustion zone, unburned zone, burnout zone, and flame front beacon zone of the flame;

[0013] Step 5: Perform statistical analysis on the parameters from Step 4 to obtain the combination quantity. In terms of combination quantity The confidence interval is used to distinguish between flame stability and flame instability, thereby determining the stability of the flame.

[0014] Furthermore, the self-luminous OH* and CH* flame images in step 1 are flame images obtained by taking pictures with a high-speed camera equipped with filters of different operating wavelengths.

[0015] Furthermore, the filter used to capture the self-emissive OH* flame image operates at a wavelength of 270-290nm or 300-320nm, while the filter used to capture the self-emissive CH* flame image operates at a wavelength of 380-400nm or 420-440nm.

[0016] Furthermore, the flame graph preprocessing in step 1 includes image filtering and denoising, image region cropping, and image bilinear interpolation scaling.

[0017] Furthermore, the unburned region mentioned in step 3 is a closed region formed by the flame front beacon zone of the self-luminous CH* flame map and its combustion area background. The minimum concavity and convexity surface of the scatter plot is extracted in the self-luminous OH* flame map using a scatter plot contour algorithm as the component intensity map of the self-luminous OH*. The unburned region is removed from the component intensity map of the self-luminous OH* to obtain the main combustion region of the flame. The combustion region obtained from the high-speed flame map is subtracted from the unburned region and the main combustion region of the flame to obtain the remaining burnout region.

[0018] Furthermore, the thermodynamic parameters in step 4 include the temperature information and non-uniformity of the combustion target area of ​​the high-speed combustion diagram; the geometric parameters include the area of ​​the main combustion zone, unburned zone, and burnout zone of the flame, and the distance to the front end of the flame front beacon; the component parameters include the relative concentration of self-luminous OH* and CH* flame diagrams; and the frequency domain parameters include the combustion scintillation frequency obtained by performing an inverse Fourier transform on the brightness information of the time-series image.

[0019] Furthermore, the steps for quantitatively monitoring flame stability are as follows:

[0020] S01: Cluster the following nine shallow feature parameters: temperature information of the combustion target area, non-uniformity, area of ​​the main combustion zone, area of ​​the unburned zone, area of ​​the burnout zone, distance of the front end of the leading edge of the flame, relative concentration of self-luminous OH* and CH* flame diagrams, and combustion scintillation frequency. Find the corresponding cluster centers and calculate the flame stability index - RSI1 with the vector to be measured to achieve flame stability classification.

[0021] S02: Divide the three types of flame images into training set, validation set and test set samples according to a certain number of components, put each sample into a 2D convolutional autoencoder neural network, train and recombine to form a new feedforward convolutional neural network, and extract the deep feature vector matrix of the three types of flame images.

[0022] S03: The extracted deep feature vector matrix is ​​classified, and flame stability and flame instability are distinguished according to step 5. The cluster centers of the two classes are calculated. The flame stability index RSI2 is calculated based on the clustering results of the test vector and the three types of flame images to achieve a quantitative expression of flame stability.

[0023] S04: Combine a certain number of time-series three types of flame images into the corresponding sample volume. Based on the determination of flame stability in step 5, set the conditions for selectively reconstructing images. By training a 3D selective convolutional autoencoder, obtain the display of signs of stable to unstable transition and the flame stability index RSI3.

[0024] S05: Input three types of flame images under the same working condition into a pre-trained convolutional autoencoder feature extraction and clustering model, and calculate the comprehensive flame stability index (RSI) of the three types of flame images by combining RSI1, RSI2 and RSI3. The stability of the flame is judged based on the final output value of the comprehensive flame stability index (RSI).

[0025] Furthermore, the training set samples comprised of the three types of flame images account for no less than 50% to 90%, the validation set samples account for 5% to 25%, and the test set samples account for 5% to 25%.

[0026] Furthermore, considering the range of the flame stability index RSI to be [0, 1], the expression is:

[0027]

[0028] Where RSI is between [0, 0.3], it indicates that the flame is extinguished;

[0029] An RSI between (0.3, 0.6) indicates an unstable flame.

[0030] An RSI value between (0.6, 1.0) indicates a stable flame.

[0031] Furthermore, both 2D convolutional autoencoder neural networks and 3D selective convolutional autoencoder neural networks consist of a convolutional encoder and a convolutional decoder. The loss function of the 2D convolutional autoencoder neural network is the sum of the similarity of learnable perceptual image patches, peak signal-to-noise ratio, and image mean square error, while the loss function of the 3D selective convolutional autoencoder neural network is the image mean square error.

[0032] Beneficial effects:

[0033] (1) Compared with the calculation speed of flame stability index of shallow features, the present invention can identify flame stability under steady-state load and rising and falling load in real time without the need for labeled images.

[0034] (2) The present invention uses three types of flame images coupled to interpret flame stability and displays the coherent structure of the image when combustion is unstable. Compared with relying solely on high-speed flame images to identify flame stability, the flame stability identification accuracy is improved by 34%.

[0035] (3) Taking into account the flame stability index with shallow characteristics, the flame stability index with deep stable load characteristics, and the flame stability index with deep rising and falling load characteristics, a comprehensive flame stability index RSI is proposed, which realizes quantitative identification of the degree of flame stability and provides guidance for advanced combustion adjustment; compared with the flame stability index that relies on shallow characteristics, the identification accuracy is improved by 23%.

[0036] (4) The intelligent model built in this invention not only has high prediction accuracy, but also has robustness in overcoming random interference. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the flame stability monitoring method of the present invention.

[0038] Figure 2 This is a flowchart illustrating the specific process of the flame stability monitoring method of the present invention.

[0039] Figure 3 These are three types of flame images under different combustion conditions according to the present invention.

[0040] Figure 4 This is a schematic diagram illustrating the principle of the inverse Abel transform of the present invention.

[0041] Figure 5 This is a schematic diagram of the "three regions and one belt" of the present invention.

[0042] Figure 6 This is a schematic diagram of the combustion pretreatment classification results of the present invention.

[0043] Figure 7 This is a schematic diagram of the 2D convolutional autoencoder network of the present invention.

[0044] Figure 8 This is a schematic diagram of the 3D selective convolutional autoencoder network of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0046] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0047] like Figure 1 , 2 As shown, a flame stability monitoring method based on image analysis and data fusion includes the following steps:

[0048] Step 1: Acquire three types of flame images under different combustion load conditions: self-luminous OH* flame image, CH* flame image, and high-speed flame image. Among them, the self-luminous OH* and CH* flame images are preprocessed by region cropping, filtering and denoising, and bilinear interpolation scaling to obtain the corresponding flame image one. The high-speed flame image is preprocessed by image region cropping, filtering and denoising, and bilinear interpolation scaling to obtain flame image two. The image region cropping, filtering and denoising, and bilinear interpolation scaling preprocessing can reduce the amount of image processing.

[0049] like Figure 3 Three types of flame images were collected under different load combustion conditions. The air-fuel ratios for the flame images were 0.9, 1.0, 1.1, 1.2, and 1.3, which correspond to different boiler loads in practice.

[0050] The self-luminescent OH* and CH* flame images in step 1 are flame images obtained by taking pictures with a high-speed camera equipped with filters of different operating wavelengths.

[0051] Specifically, the image resolution of the self-emissive OH* and CH* flame graph is M1×N1, and the number of image frames is s frames per second; the image resolution of the high-speed flame graph is M2×N2, and the number of image frames is q frames per second. For example, the image resolution of the self-emissive OH* and CH* flame graph is 1280×1024, and the number of image frames is 10 frames per second; the image resolution of the high-speed flame graph is 2252×2252, and the number of image frames is 50 frames per second.

[0052] In one embodiment, the filter used to capture a self-emissive OH* flame image operates at a wavelength of 270-290nm or 300-320nm, and the filter used to capture a self-emissive CH* flame image operates at a wavelength of 380-400nm or 420-440nm.

[0053] Step 2: Perform inverse Abel transform on the self-illuminating CH* flame image to extract the flame frontal band, and perform a scatter contour algorithm on the self-illuminating OH* flame image to extract the main combustion zone of the flame image, such as... Figure 4 As shown.

[0054] Step 3: The flame front beacon zone and flame image two form a closed unburned region. The difference between the main combustion region and flame image one yields the burnout region, thus obtaining... Figure 5 The “three zones and one area” shown are the unburned zone, the main combustion zone, the burnout zone, and the front beacon zone of the flame.

[0055] Step 4: Extract the thermodynamic parameters, geometric parameters, composition parameters, and frequency domain parameters corresponding to the main combustion zone, unburned zone, burnout zone, and flame front beacon zone ("three zones and one zone"). The thermodynamic parameters include the temperature information and non-uniformity of the combustion target area in the high-speed flame image. The geometric parameters include the area of ​​the main combustion zone, unburned zone, and burnout zone and the distance to the front end of the flame front beacon zone. The composition parameters include the relative concentration of self-luminous OH* and CH* in the flame image. The frequency domain parameters include the combustion scintillation frequency obtained by performing an inverse Fourier transform (FFT) on the brightness information of the time-series image.

[0056] The temperature information of the combustion target area is obtained using the two-color method. Since the second radiation constant in Planck's law is relatively small, the intensity of monochromatic light is proportional to temperature at relatively high temperatures. Therefore, the temperature information can be approximated by the image brightness. The non-uniformity of the combustion area represents the brightness information in the image; the area of ​​the "three zones" only represents the number of pixels in different areas. The distance from the front of the flame front to the burner outlet is taken from the number of pixels at the vertical distance from the combustion front surface to the burner end face. The relative concentrations of self-luminous OH* and CH* flame images in the component parameters are taken from the ratio of the total number of luminous points in that area to the area of ​​that area. The combustion flicker frequency is the specific value obtained by performing an inverse Fourier transform (FFT) on the brightness change over 5 to 10 seconds. The specific calculation formulas are shown in Table 1 below.

[0057] Table 1. Parameter Calculation Formulas for "Three Regions and One Belt"

[0058]

[0059] Where: W×H represents the image width and height; x i,j This represents the image brightness, ranging from [0, 255]; p i f is the power spectral density;i N1 represents the corresponding frequency; N2 represents the number of frequencies; N1 represents the number of samples. This represents the Fourier transform.

[0060] Step 5: Perform statistical analysis on the nine shallow-layer characteristic parameters from Step 4: temperature information of the combustion target area, non-uniformity, area of ​​the main combustion zone, area of ​​the unburned zone, area of ​​the burnout zone, distance of the leading edge of the front beacon, relative concentration of self-luminous OH* and CH* flame patterns, and combustion scintillation frequency. The combined amount is obtained from the statistical analysis. In terms of combination quantity The confidence interval of 95.4% is used to distinguish between flame stability and flame instability; that is, if the first 8 shallow feature parameters all fall within the 95.4% confidence interval of their respective intervals and the dominant combustion flicker frequency is between 7 and 14 Hz, the flame is considered stable; if they fall outside the 95.4% confidence interval of their respective intervals and the dominant combustion flicker frequency is not between 7 and 14 Hz, the flame is considered unstable. The classification results of flame stability determination are as follows: Figure 6 As shown, the above nine shallow feature parameters are clustered using the K-means model to find the corresponding cluster centers, and the flame stability index - RSI1 is calculated with the vector to be tested to achieve flame stability classification.

[0061] The vector to be tested is a column vector composed of the above nine shallow feature parameters, calculated based on the flame image (including the self-illuminating image) obtained by taking a picture.

[0062] Step 6: Divide the high-speed flame graph, self-illuminating OH* and CH* flame graph from Step 1 into training set, validation set and test set samples according to a certain number of components; the training set is only used for model update parameters, the validation set is used to select the optimal model, and the test set is mainly used for model generalization ability detection.

[0063] This sample is only used for data samples under steady-state load. For rising and falling load conditions, the above sample set needs to be divided into two types of sample sets according to step 5: flame stability and flame instability. The sample volume needs to be made according to the interval k over time, taking the time factor in the image into account. The target image of the stable sample is a completely black image, while the target image of the unstable sample is itself.

[0064] Step 7: Input the training set, validation set, and test set samples from Step 6 into the 2D convolutional autoencoder neural network, train the parameters of the convolutional encoder and convolutional decoder, save the parameters of the trained convolutional autoencoder network, and reassemble the convolutional encoder in the convolutional autoencoder network in sequence to form a brand-new feedforward convolutional neural network, and extract the deep feature vector matrix of the three types of flame images.

[0065] like Figure 7As shown, since shallow feature parameters often involve a large amount of computation, it is difficult to achieve real-time monitoring data. In practice, they are set to be calculated once every minute, and their values ​​are applied within that minute. Flame stability is calculated in real time using 2D convolutional autoencoders.

[0066] It should be noted that all three types of flame images are two-dimensional matrices. The model can find a feature vector matrix that best represents the information of the two-dimensional matrix from the image. This matrix contains all the basic information of the image, and this basic information can reconstruct the image with very little error compared to the original image. This indicates that this feature vector matrix can represent the two-dimensional matrix of the image.

[0067] Step 8: Use the K-means model to perform binary classification on the deep feature vector matrix extracted in Step 7. Based on the flame stability determination in Step 5, use it as the two-class distinction between flame stability and combustion instability to achieve the distinction between the two classes. Calculate the cluster centers of the two classes. Calculate the flame stability index RSI2 based on the clustering results of the test vector and the three types of flame images to achieve a quantitative expression of flame stability.

[0068] Step 9: Combine a certain number of time-series three types of flame images to form the corresponding sample volume. Based on the determination of flame stability in Step 5, set the conditions for selective image reconstruction, that is, the flame stable sample volume is reconstructed into a black and white image and the combustion unstable sample volume is reconstructed into the original image. Through 3D convolutional autoencoder parameter update, realize the display of signs of stable to unstable transition and the flame stability index RSI3, where the flame stability index RSI3 is the average flame stability index of the three types of flame images.

[0069] The specific process of model training can be found here. Figure 7 and Figure 8 If the loss function curve in the trained model is difficult to decrease, it is because the model parameters are too large, making it difficult for backpropagation to reach the previous convolutional layers. Therefore, an optimization strategy is provided here: a greedy algorithm can be used to split several convolutional networks and assemble them into several autoencoder layers, gradually training and updating the parameters to achieve initial training. The parameters are not optimal, but only locally optimal. Finally, they are reassembled into the final convolutional autoencoder neural network and retrained to find the optimal model.

[0070] Step 10: Input the three types of flame images under the same working condition into the pre-trained convolutional autoencoder feature extraction and clustering model, and calculate the comprehensive flame stability index RSI of the three types of flame images by combining RSI1, RSI2 and RSI3. The combustion stability is judged based on the final output value of the comprehensive flame stability index RSI.

[0071] This invention classifies shallow features of three types of flame images into two categories based on the confidence interval concept in fault diagnosis: flame stability and flame instability. Because image data is inherently high-dimensional and large in volume, traditional image feature extraction is inconvenient. Therefore, conventional shallow combustion feature extraction lags behind actual combustion monitoring. Thus, an autoencoder neural network is used to reduce the dimensionality of the image data and accelerate the process. Conventional shallow combustion feature calculation can still be used in the data preprocessing stage to establish datasets for flame stability and instability.

[0072] Currently, conventional flame stability monitoring is often used for monitoring under steady-state high loads, lacking guidance for combustion monitoring under low loads and variable load processes. Therefore, this invention introduces a steady-state load flame stability monitoring method and a fluctuating load flame stability monitoring method. Compared to the calculation speed of shallow feature flame stability indices, this method can identify flame stability under steady-state and fluctuating loads in real time, without requiring labeled images. The steady-state load flame stability monitoring method can monitor flame stability not only under steady-state high loads but also under low loads. The fluctuating load flame stability monitoring method is based on the reconstruction characteristics of autoencoder neural networks, incorporating the feature of selective combustion reconstruction. It only reconstructs instability factors in flame images, enabling real-time monitoring of combustion instability under fluctuating loads and providing early warning of excessive fluctuating load rates. This will be helpful in selecting the optimal fluctuating load rate curve during future deep peak shaving.

[0073] The comprehensive flame stability index combines three categories: flame stability monitoring indicators based on the aforementioned data preprocessing, steady-state load flame stability monitoring indicators, and fluctuating load flame stability monitoring indicators. Data preprocessing takes approximately one minute, during which the output flame stability index continues the previous stability index. Based on the coal feed rate versus time curve, comprehensive evaluation index coefficients are established for steady-state load and fluctuating load. Among the two types of comprehensive evaluation index coefficients, the weight of the preprocessed flame stability monitoring indicators is smallest under fluctuating load, while the weight of the fluctuating load flame stability monitoring indicators is smallest under steady-state load. This comprehensive flame stability evaluation method has a wide range of applications, not limited to monitoring low-load steady-state and fluctuating loads. This method can also quantitatively display the flame stability coefficient, providing guidance in practical industrial applications.

[0074] like Figure 4 As shown, in the above steps, step 2 uses the inverse Abel transform to extract the flame front beacon band from the flame image. The principle is that the projected three-dimensional axisymmetric flame image is projected onto a two-dimensional plane. The original flame image is then calculated using the projection value, which is obtained by inverting the chord integral of the flame image, thus yielding the position and shape of the flame front. The actual brightness f(r) at position r in the top view of the self-illuminating CH* flame image is:

[0075]

[0076] Where I(x) represents the cumulative brightness value in the self-luminous CH* flame diagram, which is approximated by a cubic function, i.e., I(x) = a + bx + cx. 2 +dx 3 R is the distance from the combustion edge to the combustion center in the top view of the self-illuminating CH* flame diagram, and r is the distance from (x,y) to the combustion center in the top view.

[0077] In the above steps, the unburned region mentioned in step 3 is a closed region formed by the flame front beacon zone of the self-luminous CH* flame map and its combustion area background. The minimum concavity and convexity surface of the scatter plot is extracted in the self-luminous OH* flame map using a scatter plot algorithm as the component intensity map of the self-luminous OH*. The unburned region is removed from the component intensity map of the self-luminous OH* to obtain the main combustion region of the flame. The combustion region obtained from the high-speed flame map is subtracted from the unburned region and the main combustion region of the flame to obtain the remaining burnout region.

[0078] This invention utilizes a scatter profile algorithm to determine high-temperature regions in self-luminous images and calculates the number of self-luminous scatter points within these regions, thereby achieving quantitative analysis of the intensity of self-luminous components and increasing the accuracy of flame stability classification during data preprocessing.

[0079] In the above steps, the combination quantity in step 5 The expression is:

[0080]

[0081] Where P is the principal component vector of the trained data, Λ is the diagonal matrix of the ordered eigenvalues ​​of the corresponding principal component vector P, and χ 2 δ represents the 95.4% confidence interval for the squared prediction error. 2 denoted as the 95.4% confidence interval of the Hotling statistic, and x represents the shallow feature column vector in the image to be tested.

[0082] The flame stability index RSI1 has a numerical range of [0, 1], and its expression is:

[0083]

[0084] In the formula, the eigenvector matrix X m×n =[x1,x2,…,x n ], where x i =[a 1i ,a 2i ,…,a mi ] T , ||x i -x j The || operator represents a vector (x)i -x j ) and transpose vector (x) i -x j ) T The square root of the product;

[0085] This represents the i-th column vector of the cluster center vector of the deep characteristic vector matrix of stable combustion. represents the i-th column vector of the cluster center vector of the deep characteristic vector matrix of unstable combustion, where the cluster center vector of the deep characteristic vector matrix of combustion is the optimal center vector obtained by K-means calculation.

[0086] In the above steps, the numerical range of the flame stability index RSI2 in step 8 is [0, 1], and the expression is:

[0087]

[0088]

[0089] In the formula, The second indicator represents the flame stability of a high-speed flame diagram;

[0090] The second indicator of flame stability in a self-luminous OH* flame diagram;

[0091] The second indicator represents the flame stability index of the self-luminous CH* flame diagram.

[0092] Flame stability and flame instability are classified into two categories based on feature vectors. The cluster center vectors of the two categories are calculated to achieve quantitative calculation of the flame image under test. This calculation process is completed in milliseconds, enabling real-time monitoring of flame stability.

[0093] In the above steps, the numerical range of the flame stability index RSI3 in step 9 is [0, 1], and the expression is:

[0094]

[0095] In the formula, This represents the pixel value of the reconstructed image at (i,j), which ranges from [0, 255].

[0096] The pixel value at (i,j) in the original image representing the three types of flame images ranges from [0, 255].

[0097] The k value indicates the flame image type: k=1 indicates a high-speed flame image, k=2 indicates a self-luminous OH* flame image, and k=3 indicates a self-luminous CH* flame image.

[0098] This invention provides a quantitative description of flame stability based on reconstructed unstable flame images, reflecting the degree of instability during combustion in real time. By re-marking the regions of instability in the three types of flame images to their corresponding locations in the image, it provides effective indicators of instability and displays the coherent structure appearing in the image when combustion is unstable, facilitating operation and adjustment by personnel. Compared to relying solely on high-speed flame images to identify flame stability, the accuracy of flame stability identification is improved by 34%.

[0099] In the above steps, in step 6, the training set samples for classifying the three types of flame images account for no less than 50% to 90%, the validation set samples account for 5% to 25%, and the test set samples account for 5% to 25%. For example, the training set samples for classifying the three types of flame images account for 80%, the validation set samples account for 10%, and the test set samples account for 10%.

[0100] In the above steps, the 2D convolutional autoencoder neural network and the 3D selective convolutional autoencoder neural network in step 7 are composed of a convolutional encoder and a convolutional decoder. They are internally set with upsampling, convolution, batch normalization layer, pooling layer and deconvolution layer. The loss function of the 2D convolutional autoencoder neural network is the sum of the similarity of learnable perceptual image patches, peak signal-to-noise ratio and image mean square error. The loss function of the 3D selective convolutional autoencoder neural network is the image mean square error.

[0101] Wherein, the loss function L MSE The expression is:

[0102]

[0103] In the formula, and These represent the pixel values ​​at (i,j) of the reconstructed and original images of the three types of flame images with a size of W×H, respectively.

[0104] Loss function L PSNR The expression is:

[0105]

[0106] In the formula, C represents the maximum grayscale value that the image can reach;

[0107] Loss function L LPIPS The expression is:

[0108]

[0109] In the formula, ω l Let H be the weight matrix of the l-th channel. l With W lThese represent the size of the flame image in the l-th channel. and This represents the pixel value at (i,j) of the original and reconstructed images of the l-th channel.

[0110] The loss function of a convolutional autoencoder neural network is divided according to its different functions. For extracting deep features from before and after flame images, the reconstructed image should be as close as possible to the original image; the more accurate the mid-to-deep features, the more accurate the loss function. Therefore, in addition to considering peak signal-to-noise ratio and image mean square error, learnable perceptual image patch similarity is also considered. For selectively reconstructing unstable images, strong generalization ability is required, so high image accuracy is not required. Therefore, the image mean square error is used as the loss function.

[0111] In the above steps, step 10 comprehensively judges that the flame stability index RSI is within the range of [0, 1], and the expression is:

[0112]

[0113] Where RSI is between [0, 0.3], it indicates that the flame is extinguished;

[0114] An RSI between (0.3, 0.6) indicates an unstable flame.

[0115] An RSI between (0.6, 1.0) indicates a stable flame.

[0116] When the flame goes out, the system reports a fault and requires emergency control; when the flame is unstable, continuous monitoring and appropriate control are required; when the flame is stable, it maintains autonomous operation without intervention.

[0117] Finally, the test sample data achieved quantitative identification of flame stability, providing guidance for advanced combustion adjustments. Compared with flame stability indicators that rely on shallow features, the identification accuracy was improved by 23%. Compared with traditional combustion monitoring, it has real-time, qualitative and quantitative flame images of flame stability. Combined with self-luminous OH* flame maps and CH* flame maps, it improved the flame stability monitoring effect and visualized areas of instability.

[0118] In summary, considering the flame stability indices based on shallow characteristics, deep stable load characteristics, and deep rising and falling load characteristics, a comprehensive flame stability index (RSI) is proposed. This index enables quantitative identification of flame stability and provides guidance for advanced combustion adjustments.

[0119] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A flame stability monitoring method based on image analysis and data fusion, characterized in that, Includes the following steps: Step 1: Acquire three types of flame images under different combustion load conditions: self-luminous Flame diagram Flame diagrams and high-speed flame diagrams, including self-illuminating flame diagrams. and Flame image 1 is obtained by preprocessing the flame image, and flame image 2 is obtained by preprocessing the high-speed flame image. Step 2: For self-illuminating... The flame image is subjected to inverse Abel transform to extract the flame front beacon and self-illumination. The main combustion region of the flame is extracted from the flame image using a scatter contour algorithm. Step 3: The flame front beacon zone and the second flame image form a closed unburned region, and the main combustion region of the flame is subtracted from the first flame image to obtain the burnout region; Step 4: Extract the thermodynamic parameters, geometric parameters, composition parameters, and frequency domain parameters corresponding to the main combustion zone, unburned zone, burnout zone, and flame front beacon zone of the flame; Step 5: Perform statistical analysis on the parameters from Step 4 to obtain the combination quantity. , in combination quantity The confidence interval is used to distinguish between flame stability and flame instability, thereby determining the stability of the flame; The steps for quantitatively monitoring flame stability are as follows: S01: Information on the temperature, non-uniformity, area of ​​the main combustion zone, area of ​​the unburned zone, area of ​​the burnout zone, distance from the front edge of the flame front, and self-luminescence of the combustion target area. and Nine shallow feature parameters, including the relative concentration and combustion flash frequency of the flame diagram, are clustered, and the corresponding cluster centers are found. The flame stability index - RSI1 is calculated with the vector to be tested to achieve flame stability classification. S02: Divide the three types of flame images into training set, validation set and test set samples according to a certain number of components, put each sample into a 2D convolutional autoencoder neural network, train and recombine to form a new feedforward convolutional neural network, and extract the deep feature vector matrix of the three types of flame images. S03: The extracted deep feature vector matrix is ​​classified, and flame stability and flame instability are distinguished according to step 5. The cluster centers of the two classes are calculated. The flame stability index RSI2 is calculated based on the clustering results of the test vector and the three types of flame images to achieve a quantitative expression of flame stability. S04: Combine a certain number of time-series three types of flame images into the corresponding sample volume. Based on the determination of flame stability in step 5, set the conditions for selectively reconstructing images. By training a 3D selective convolutional autoencoder, obtain the display of signs of stable to unstable transition and the flame stability index RSI3. S05: Input three types of flame images under the same working condition into a pre-trained convolutional autoencoder feature extraction and clustering model, and calculate the comprehensive flame stability index (RSI) of the three types of flame images by combining RSI1, RSI2 and RSI3. The stability of the flame is judged based on the final output value of the comprehensive flame stability index (RSI).

2. The flame stability monitoring method according to claim 1, characterized in that, Self-illumination in step 1 and A flame image is a picture of a flame taken by a high-speed camera equipped with filters of different operating wavelengths.

3. The flame stability monitoring method according to claim 2, characterized in that, Self-illuminating photography The flame imager is equipped with filters that operate at wavelengths of 270-290nm or 300-320nm, capturing self-illuminating images. The filters used in flame diagrams operate at wavelengths of 380-400nm or 420-440nm.

4. The flame stability monitoring method according to claim 1, characterized in that, The flame graph preprocessing in step 1 includes image filtering and denoising, image region cropping, and image bilinear interpolation scaling.

5. The flame stability monitoring method according to claim 1, characterized in that, The unburned area mentioned in step 3 is composed of self-luminous material. The closed area formed by the flame front band and the background of the burning area in the flame diagram is self-luminous. The scatter plot algorithm is used to extract the minimum concavity / convexity surface of the scatter plot as the self-illumination in the flame plot. Component intensity diagram, self-luminescence The unburned region is removed from the component intensity map to obtain the main combustion region of the flame. The combustion region obtained from the high-speed flame map is subtracted from the unburned region and the main combustion region of the flame to obtain the remaining burnout region.

6. The flame stability monitoring method according to claim 1, characterized in that, The thermodynamic parameters in step 4 include the temperature information and non-uniformity of the combustion target area in the high-speed combustion diagram; the geometric parameters include the area of ​​the main combustion zone, unburned zone, and burnout zone, and the distance to the front end of the flame front beacon; and the composition parameters include self-luminescence. and The relative concentration of the flame image, and the frequency domain parameters include the combustion flicker frequency obtained by inverse Fourier transform of the brightness information of the time-series image.

7. The flame stability monitoring method according to claim 1, characterized in that, The training set samples comprised of three types of flame images account for no less than 50% to 90%, the validation set samples account for 5% to 25%, and the test set samples account for 5% to 25%.

8. The flame stability monitoring method according to claim 1, characterized in that, The range of the flame stability index RSI is comprehensively judged within The expression is: Where RSI is between [0, 0.3], it indicates that the flame is extinguished; An RSI between (0.3, 0.6) indicates an unstable flame. An RSI value between (0.6, 1.0) indicates a stable flame.

9. The flame stability monitoring method according to claim 1, characterized in that, 2D convolutional autoencoder neural networks and 3D selective convolutional autoencoder neural networks are composed of convolutional encoders and convolutional decoders. The loss function of 2D convolutional autoencoder neural networks is the sum of the similarity of learnable perceptual image patches, peak signal-to-noise ratio, and image mean square error. The loss function of 3D selective convolutional autoencoder neural networks is the image mean square error.

Images